🔍 Read the full analysis: Granite Speech 5.0: A Faster Approach To Accurate Transcription And Attribution on ThorstenMeyerAI.com
TL;DR
IBM announced one new model and revisions to two existing models in its Granite Speech collection. The releases target multilingual transcription and translation, faster transcription, and improved speaker attribution and timestamps; the performance figures cited are IBM’s own evaluations and do not establish results across all real-world audio.
IBM has announced three updates to its Granite Speech collection: a new multilingual 2-billion-parameter model, a faster multilingual revision of its 470-million-parameter Turbo CTC model, and updated weights for a model focused on speaker attribution and timestamps. IBM says the releases target different speech workloads, including translation, high-volume transcription and long recordings with multiple speakers; the reported performance gains come from evaluations cited in the original analysis.
The new granite-speech-5.0-2b replaces the 4.1 2B model at the same parameter count. IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese. It also supports translation to and from English for those languages, along with translation from English into Italian and Mandarin. According to IBM, training used 390,000 hours of public audio plus synthetic data for Japanese speech recognition, keyword biasing and translation.
IBM says the model is better than its predecessor at handling noise, reverberation and distant microphones, and can accept a prompt specifying the input language. Users can provide keyword lists of up to 2,000 terms to guide recognition or translation. The company reports capitalization F1 scores from 89.7 to 99.5 across its test sets when formatted output is requested. Those scores are reported as a range; the supplied material does not give a breakdown by test set.
The revised granite-speech-5.0-470m-turboctc adds Japanese, Spanish, German, French, Italian and Portuguese transcription to English, as well as native punctuation and capitalization. IBM reports 19,870 RTFx on the Open ASR Leaderboard, up from about 12,600 at the model’s August launch, while saying accuracy stayed at the same level. RTFx is a speed measure, and the stated figures are IBM’s reported results, not an independent deployment assessment.
For granite-speech-4.1-2b-plus, IBM reports a 24% reduction in word error rate on long-form test sets, a 7% reduction in two speaker-attribution error measures, and more than a halving of average timestamp shift across speech. The company says the update keeps the architecture, prompt format and API unchanged. All three models are released under the Apache 2.0 license and have native support in Transformers; IBM also lists vLLM, llama.cpp and mlx-audio support for the new 5.0 2B model.
Choosing Models for Speech Workloads
The releases offer organizations different choices for different audio tasks, rather than a single model positioned to handle every requirement. The 5.0 2B model combines multilingual transcription with translation, Turbo CTC emphasizes speed across more languages, and the 4.1 2B Plus revision targets long-form recordings where speaker labels and timing can matter.
Those distinctions can affect work after audio is converted to text. More accurate timestamps may help align captions or locate passages in recordings, while speaker attribution can make meetings and support calls easier to review. Keyword lists may help with names, acronyms and specialized vocabulary. These are plausible uses of the stated features, but the supplied material does not quantify improvements in customer deployments.
IBM’s results give prospective users measures to investigate, but public benchmark performance is not a guarantee for every accent, language variety, microphone, or business workflow. Organizations would still need to test the models against representative audio and account for transcript review, infrastructure and deployment costs.
professional speech transcription software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Granite Models Differ
Granite Speech is IBM’s collection of models for speech recognition and related audio tasks. IBM introduced the 470-million-parameter Turbo CTC model in August as an English-only encoder designed for speed. The October revision retains its size while adding six transcription languages and features for punctuation and capitalization.
The new 5.0 2B model succeeds the 4.1 2B version and is positioned for multilingual transcription and translation. The updated 4.1 2B Plus is aimed at speaker attribution, timing and longer recordings. That division reflects different evaluation needs: word error rate measures transcription errors, RTFx measures processing speed, and speaker or timestamp measures assess other aspects of a transcript.
IBM says the 5.0 2B model ranked second among open-weight ASR models on average English word error rate, reporting 5.06%. The company calls this result unofficial and says it averages public English short-form test sets on the Open ASR Leaderboard as of October 15, 2026. Its Turbo CTC speed comparison uses nine public English short-form test sets and is dated October 2026. The figures do not describe every language or recording condition.
“The model is the successor to granite-speech-4.1-2b at the same parameter count.”
— IBM, in its Granite Speech announcement
Limits of the Published Evaluations
The announcement’s reported results have not been independently validated in the supplied material. IBM’s short-form leaderboard figures do not establish performance on all long recordings, languages, accents, overlapping speech or noisy environments. The source material also does not provide enough detail to judge how the models perform on specialized vocabulary or on audio from specific business settings.
Several comparison details remain incomplete. IBM reports Turbo CTC’s speed increase while describing accuracy as unchanged, but the supplied material does not include the full benchmark table or specify the hardware and deployment conditions behind the RTFx figures. The capitalization F1 range for the 5.0 2B model is not broken down by test set. Nor does the announcement quantify effects in customer deployments, such as changes in review time or transcription costs.
The models’ Apache 2.0 licensing and listed software support are stated, but the material does not set out the full compute requirements or operational costs of using them. Teams will need to consult the model cards and evaluate the systems on their own audio before drawing conclusions about suitability.
Testing and Deployment Decisions
IBM directs users to the models’ model cards for evaluation results, architecture and training information. Those documents, along with tests on recordings representative of a team’s actual use, are the next sources for assessing language coverage, accuracy, speed and resource needs.
The announcement does not specify a further release date or a separate validation milestone. Organizations considering adoption can compare the models against their current transcription process, including how often people must correct transcripts and speaker labels. IBM also recommends pairing the models with Granite Guardian for production deployments that need additional risk detection; the supplied material does not quantify the effect of that pairing.
Key Questions
What did IBM announce?
IBM announced one new Granite Speech model and revisions to two existing models, targeting multilingual transcription and translation, faster transcription, and improved speaker and timestamp measures.
Which languages does the new 5.0 2B model support?
IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese. It supports translation to and from English for those languages, and translation from English into Italian and Mandarin.
How fast is the revised Turbo CTC model?
IBM reports a score of 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at its August launch. The supplied information does not specify the hardware or deployment conditions behind the reported figures.
Are the reported improvements independently confirmed?
The supplied material presents them as IBM-reported evaluation results and provides no independent validation. Benchmark results also do not establish performance for every accent, recording environment or customer workflow.
What license do the models use?
IBM says all three releases are available under the Apache 2.0 license. The company lists native Transformers support for all three and additional software support for the new 5.0 2B model.
Primary source: Hugging Face · via ThorstenMeyerAI.com
